Solving the "Long-Tail" problem in cardiac arrhythmia detection with Stratified Deep Learning.
Utilizes stratified resampling to ensure equal representation for all arrhythmia types, preventing the model from ignoring rare, life-threatening "outlier" events.
Combines heterogeneous data sources (MIT-BIH + PTB) with adaptive training to achieve state-of-the-art 97.53% accuracy on held-out test sets.
Employs non-linear time-warping and Gaussian noise injection to force the model to learn robust morphological features rather than memorizing templates.
Traditional ECG models often achieve high accuracy only by overfitting to the majority class (Normal Sinus Rhythm) while failing to detect rare, critical arrhythmias like Supraventricular Ectopic Beats (S-class) or Fusion Beats (F-class).
Vitalis solves this by explicitly engineering against this "Long-Tail" effect. Our Stratified Deep Learning framework demonstrates a 67% relative improvement in F1-macro score (0.9017 vs 0.54 baseline), establishing a reliable foundation for real-world automated rhythm analysis.
Developed by researchers from ENMED – School of Engineering Medicine, Texas A&M University and Houston Methodist Hospital.